3D Convolutional Neural Networks for Brain Tumor Segmentation

Abstract

Brain Tumor segmentation aims to separating the different tumor tissues (solid or active tumor, edema, and necrosis) from normal brain tissue: gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF). In brain tumor studies, the existence of abnormal tissues may be easily detectable most of the time. However, accurate and reproducible segmentation and characterization of abnormalities are not straightforward. Even for experienced doctors, precise detection of tumor in MR image makes a long business. In recently years, Convolutional Neural Networks (CNNs) often obtain incredible results in problems to extract information from complex and high-dimensional inputs, for which useful features are not obvious to be identified from the structured design. In this thesis, we propose a new model for brain tumor segmentation in Magnetic Resonance Imaging (MRI) by adapting multiple dimension feature combination. First, we preprocessed original MRI meta-image in several way because MRI meta-image has much information not only brain tissue and tumor but also unnecessary noise, skull, biased intensity. Second, we built two dimensional multi-channel non-pooling convolutional neural networks. It is not common model because almost convolutional neural networks models have focused on natural image classification. But in medical image processing, low level accurate segmentation is important than high level abstract. Finally, we built a three-dimensional convolutional neural networks by combination of 2 dimensional model data and compare its result with 2 dimensional CNNs result. The method was evaluated on BRATS 2013 (Brain Tumor Segmentation 2013) data set. This data set provided via the Virtual Skeleton Database (VSD). Experiment results on respective datasets show the proposed algorithm works successfully and combination of coordinate plane information improve performance in three-dimensional segmentation problem